To get an edit instead of a rewrite, tell the AI the edit level you want, list what it must leave alone, and ask it to report its changes rather than silently replace your text. "Improve this" gives the model permission to do anything, and "anything" usually means its own version of your paragraph.
Below is a measured example (one light edit versus one rewrite of the same paragraph), a table of edit levels, a copy-ready prompt, and a short routine for editing a whole chapter without losing your voice.
Why does AI rewrite when I only asked it to improve something?
Because "improve" doesn't tell it where to stop, so it optimizes the whole passage. A language model produces the most likely good-looking answer to your request. If the request is vague, the safest-looking answer is a fresh, polished paragraph. It isn't ignoring you. You haven't given it a boundary.
Anthropic's prompting guidance makes a related point: current models follow the verbs you use quite literally. Ask whether it "can suggest changes" and you may get suggestions only; tell it to change something and it will. So choose the verb on purpose: "mark", "list", "fix", "tighten" or "rewrite" each mean something different to the model.
Instructions can also fail to stick on their own terms. The IFEval benchmark (Zhou et al., 2023) tests models on about 500 prompts built from 25 kinds of verifiable instructions, such as "write in more than 400 words" or "mention a keyword at least three times". The point of the design is that you can check compliance mechanically. You can borrow that idea: give edit instructions you can verify, like "do not change the word count by more than 5%" or "keep every heading exactly as written".
What does a light edit look like compared with a rewrite?
A light edit changes a handful of words and keeps the author's sentences; a rewrite replaces most of them. Here is a rough paragraph I wrote for this post (61 words):
I started writing my first ebook in march and honestly I did not know what I was doing. Their were so many tools and I tried alot of them. In the end the thing that worked was just a simple routine, I wrote for 30 minutes every morning before work, and by june I had a first draft of 40 pages.
Here is a light copyedit. It fixes "march", "Their", "alot", the comma splice and "june", and nothing else:
I started writing my first ebook in March and honestly I did not know what I was doing. There were so many tools and I tried a lot of them. In the end the thing that worked was just a simple routine: I wrote for 30 minutes every morning before work, and by June I had a first draft of 40 pages.
And here is what "improve this" often returns:
Last March, I began my first ebook with no clear plan and a dizzying array of tools to choose from. What ultimately made the difference was consistency: thirty focused minutes each morning before work. By June, those small sessions had added up to a 40-page draft.
Both are plausible edits, and I wrote both by hand as illustrations (they are not transcripts from a particular model). I then compared them word by word with a standard sequence-matching script:
| Version | Words | Original words kept | Share kept |
|---|---|---|---|
| Original | 61 | 61 | 100% |
| Light copyedit | 62 | 56 | 92% |
| "Improved" rewrite | 46 | 11 | 18% |
The rewrite is smoother, but it also dropped "honestly I did not know what I was doing", the line that sounded like a person. It cut the paragraph by a quarter and added phrases ("dizzying array", "ultimately", "small sessions had added up") that you'd see in a thousand other AI-polished intros. If you want that voice back, see how to make AI writing sound less like AI.
Which edit level should I ask for?
Ask for the lowest level that solves your problem. The labels below are common editorial working definitions, and the "typical share of words kept" column is my own rule of thumb from examples like the one above, not a published benchmark.
| Edit level | What it may change | Typical share of words kept |
|---|---|---|
| Proofread | Typos, punctuation, capitalization | 95-100% |
| Copyedit | Grammar, consistency, usage, repeated words | 90-97% |
| Line edit | Sentence rhythm, clarity, wordiness | 70-90% |
| Developmental edit | Order, gaps, structure (as comments, not new text) | Not measured by words |
| Rewrite | Everything | Under 50% |
For developmental feedback, ask for comments only. A model that "fixes" structure by rewriting every paragraph hands you a new draft, not feedback.
What should an edit prompt contain?
Five things: the edit level, a protect list, the output format, a cap, and the text. Here is a template you can paste and adjust:
Edit level: copyedit only. Fix spelling, grammar, punctuation and repeated words.
Protect: my wording and sentence order, all headings, every number and name, and anything in quotation marks.
Do not: add new sentences, change the tone, or merge paragraphs.
Output: the corrected text first, then a numbered list of every change as "old -> new".
Cap: if you think a sentence needs more than a grammar fix, leave it and add it to a list called "Suggestions" at the end.
TEXT:
[paste your section here]
Three details do most of the work. The protect list sets what's off limits. The "old -> new" list lets you verify the edit in a minute instead of re-reading the whole page. And the "Suggestions" bucket gives the model somewhere to put its urge to rewrite, so it doesn't leak into your paragraph. Anthropic's docs also recommend explaining the reason behind an instruction ("this is my voice, readers follow it"), and a sentence of reasoning helps the model generalize the rule to cases you didn't list.
How do I edit a whole chapter without losing my voice?
Edit in passes of one job each, and check the output between passes. This is the same generate, review, refine pattern that both Anthropic and Google recommend for complex work: split it into separate steps so the output of one becomes the input of the next, and you can inspect each step. Google's Gemini prompt-design guide calls this chaining prompts. A research study backs the idea in a small way: in a 20-person user study of "AI Chains" (Wu et al., 2021), breaking a task into connected steps improved the quality of the result and gave users more transparency and control.
- Pass 1, mechanics. Copyedit with the template above. Accept or reject the numbered changes.
- Pass 2, clarity. Ask it to flag the three muddiest sentences per page and explain why, without rewriting them. Rewrite them yourself, or ask for two options of each and pick.
- Pass 3, trims. Ask for cuts only: "mark up to 10 phrases that can be deleted without changing meaning." You delete them.
- Pass 4, read aloud. This one is yours. Read the chapter out loud and fix anything you'd never say.
The same idea is built into Interactive eBook Builder: its Agent C Polisher runs separate passes for clarity, voice, pacing, layout, interactivity and a final QA check, and the instructions behind each pass are editable, so "do not rewrite my sentences" can be a standing rule instead of something you retype. Whether you use a tool or a chat window, the principle holds: one job per pass.
If you want the bigger picture of splitting a book into stages, prompt chaining for beginners walks through planning, drafting and editing as three separate steps, and our guide to planning, drafting and editing an ebook with AI shows how those steps fit a real production schedule.
What if it rewrites anyway?
Reject the output, restate the boundary, and give it a smaller piece. Long inputs make every instruction compete with the text, so paste one section at a time instead of the whole manuscript. That is also how Interactive eBook Builder is laid out: Agent A writes a Chapter Plan, and drafting and polishing then happen chapter by chapter, so no single request has to carry the whole book. If the result still drifts, paste your original and the edited version back and ask for "only the changes that fix an error, as a list." Our guide on what to do when AI gets it wrong covers the full recovery routine, and how long a prompt should be explains why short, focused edit prompts beat sprawling ones.
One caution on the "ask it to double-check itself" habit. A 2023 study (Huang et al.) found that large language models struggle to correct their own reasoning without outside feedback, and sometimes get worse after trying. For prose edits, you are the outside feedback. Read the change list.
Before you hit send
- Did you name the edit level (proofread, copyedit, line edit)?
- Did you list what to protect: voice, facts, headings, quotes?
- Did you ask for a change list in a format you can accept or reject?
- Is the text one section, not the whole book?
- Did you compare word counts afterward?
The takeaway: an edit prompt is a set of limits. Without them, you're asking for a rewrite.